Profile by name
ttperf relu maps an operation to its TT-Metal test and runs it. No path spelunking.
TT-Metal kernel profiler · CLI
Profile a Tenstorrent kernel
in one command.
ttperf runs TT-Metal's profiler, parses the result CSV, and
prints the total device kernel duration — the one number
you actually came for.
$ █
Auto-generated profile name: temp_test_add
Running test...
============================================================
TEST SUMMARY
============================================================
Test: add
Status: PASSED
Configuration: shape=(1, 1, 32, 32), dtype=bfloat16, layout=tile
CSV Path: /tmp/profile_results.csv
DEVICE KERNEL DURATION [ns] total:
0.00 ns
============================================================
0 ns
The measurement, not the ritual
Without ttperf
- Locate
tools/tracy/profile_this.py - Hand-write the
pytestinvocation - Wait, then hunt for the generated CSV path in the logs
- Open the CSV, find the right column
- Sum
DEVICE KERNEL DURATIONby hand
With ttperf
$ ttperf add
Profile name, CSV discovery, parsing, and the total — handled. You read one number.
What it does
Real CSV parsing
Finds the profiler's output CSV, verifies it exists, and sums the device kernel duration with pandas.
Shape it yourself
Override tensor --shape, --dtype, --layout, and DRAM/L1 memory per run.
Config files
Set defaults in ~/.ttperf.yaml or ./.ttperf.yaml. CLI flags always win.
CI-friendly
--quiet drops decorative output; --verbose turns on debug logging. Exit codes behave.
Find the path itself
Resolves your tt-metal root from PYTHONPATH or by walking up the working directory.
parse → run → read
The pipeline
- parseRead args and config; figure out profile name + test path or operation.
- mapResolve an operation name to its
TestEltwiseOperationsmethod. - runInvoke TT-Metal's tracy profiler over
pytest, streaming output. - extractScrape the generated CSV path from the log and confirm it's on disk.
- sumLoad the CSV and total
DEVICE KERNEL DURATION [ns]. - reportPrint the summary; optionally copy the CSV to
--output-dir.
Every supported operation
Pulled from the package's own test suite — the same
list as ttperf --list-ops. Click any chip to copy its command.
No operations match that filter.
ship it
Install
From PyPI
First run
Requires a Tenstorrent TT-Metal environment, Python 3.8+, and pytest. PyYAML is optional, for config files.